Research map / Large language models / Alignment and preferences
LifeAlign: Lifelong Alignment for Large Language Models with Memory-Augmented Focalized Preference Optimization
Junsong Li, Jie Zhou, Bihao Zhan, Yutao Yang, Qianjun Pan, Shilian Chen, Tianyu Huai, Xin Li, Qin Chen, Liang He
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SLMC (memory consolidation) on top of FPO (ablation) | LifeAlign · Lifelong alignment (BLEU-4) | BWT: -3.68 -> 0.02 | Improved |
| add FPO on top of SLMC (ablation) | LifeAlign · Lifelong alignment (BLEU-4) | BWT: -1.39 -> 0.02 | Improved |
| add SLMC alone to plain DPO baseline (ablation) | DPO with replay · Lifelong alignment (BLEU-4) | BWT: -7.67 -> -1.39 | Improved |
| add FPO alone to plain DPO baseline (ablation) | DPO with replay · Lifelong alignment (BLEU-4) | BWT: -7.67 -> -3.68 | Improved |
| add FPO and SLMC (ablation) | DPO with replay · Lifelong alignment (LLM-Judge) | Last: 49.11 -> 57.42 | Improved |
| set SLMC projection weight lambda to 0.5 (vs extremes) (ablation) | SLMC · Lifelong alignment | performance | Improved |
| set denoising threshold theta to 0.9 (ablation) | SLMC · Lifelong alignment | performance | Improved |
| use LifeAlign instead of CPPO on LLaMA-3.1-8B | LLaMA-3.1-8B-Instruct · Lifelong alignment (LLM-Judge) | BWT: -7.77 -> 2.10 | Improved |
| use LifeAlign across reverse task order (ablation) | LifeAlign · Lifelong alignment (reverse order, LLM-Judge) | BWT: 0.51 vs CPPO -5.64, ER -3.41 | Improved |
| use LifeAlign instead of L2P | L2P · Lifelong alignment 6-task benchmark (AVG) | BWT: BWT 1.85 (L2P) vs 0.91; Last 18.81 vs 37.67 | Mixed |
Closest papers by meaning
- Leveraging robust optimization for llm alignment under distribution shifts (NeurIPS 2025 · similarity 0.91)
- Understanding the Learning Dynamics of Alignment with Human Feedback (ICML 2024 · similarity 0.90)
- ComPO: Preference Alignment via Comparison Oracles (NeurIPS 2025 · similarity 0.90)
- Panacea: Pareto Alignment via Preference Adaptation for LLMs (NeurIPS 2024 · similarity 0.90)
- Clean First, Align Later: Benchmarking Preference Data Cleaning for Reliable LLM Alignment (NeurIPS 2025 · similarity 0.90)
- Spread Preference Annotation: Direct Preference Judgment for Efficient LLM Alignment (ICLR 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
